Discovery Through Constraints: Imposing Constraints on Autoencoders for Data Representation and Dictionary Learning

This article reviews data representation and latent feature extraction using several types of constrained AEs. In addition to sparsity, nonnegative RFs can be enforced during learning to enhance the interpretability of data and their additive properties where applicable. Comparisons are made between the computing of matrix Φ with quasi-basis vectors as columns containing dictionary terms and unsupervised learning of a cascade of two RFs being rows of encoding matrix W
₁
and columns of the decoding matrix W
₂
.